ORDINAL XGBOOST FOR MULTICLASS NUTRITIONAL STATUS CLASSIFICATION WITH IMBALANCED DATA

Abstract

Stunting is a major global health problem that affects children’s physical growth and cognitive development, particularly in developing countries. The classification of toddlers’ nutritional status to detect stunting risk often faces two primary challenges: the ordinal nature of the labels and the imbalance in class distribution, where minority classes such as stunted and tall are much smaller than the majority class (normal). This study aims to develop an Ordinal Extreme Gradient Boosting (Ordinal XGBoost) method using a Binary Decomposition approach to classify toddlers’ nutritional status in imbalanced ordinal data. Secondary data from 100 respondents were analyzed, with 80% allocatedfor training and 20% for testing. The Binary Decomposition approach transforms the three-class ordinal classification problem into two binary subproblems, each trained using XGBoost with weighted logistic loss to address class imbalance. Model performance was evaluated using four key metrics: accuracy, ordinal Mean Absolute Error (MAE), Quadratic Weighted Kappa (QWK), and Macro-F1. Results showed an accuracy of 70%, ordinal MAE of 0.30, QWK of 0.45, and Macro-F1 of 0.53. The MAE and QWK values indicate the model’s ability to preserve class order while reducing large prediction jumps, although performance on minority classes remains limited. These findings suggest that the proposed approach is effective for imbalanced ordinal data and has potential applications in toddler nutritional status monitoring systems.